Research map / Reinforcement learning / Visual and representation RL
Policy Design in Long-run Welfare Dynamics
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Shows, in a stochastic model of welfare dynamics with decay and intervention returns, that Rawlsian (min-U, max-g) policies achieve higher long-run average welfare than utilitarian policies under survival and Matthew-effect conditions, with a reversal under a ruin condition.
Problem
Comparing Rawlsian and utilitarian social policies on long-run social welfare, given that policies that look suboptimal in the short run may yield larger long-term benefits.
Approach
Extends an attention-allocation model from industrial project management to welfare dynamics with individual-specific return and decay functions, proves closed-form long-run growth rates for Rawlsian and utilitarian policies, and gives a survival condition under which Rawlsian policies dominate (and a ruin condition under which utilitarian policies dominate), using a Lundberg-type inequality for s
Key results
- Under survival and Matthew-effect assumptions, Rawlsian min-U and max-g achieve long-run social welfare (M/N)f+ - ((N-M)/N)g-, exceeding utilitarian (M/N)f+ - ((N-M)/N)g+ almost surely.
- In SIPP-based simulations (206 individuals), Rawlsian policies have lower short-term welfare than utilitarian ones but overtake them after a few hundred timesteps.
- Under a ruin condition, the comparison reverses and utilitarian policies achieve better long-run welfare.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Rawlsian min-U/max-g instead of utilitarian max-U/max-f/max-fg under survival, Matthew-effect, and uniform-boundedness conditions | Welfare policy · Long-run welfare under survival condition | long-term social welfare | Helped |
| Use utilitarian policies instead of Rawlsian policies under the ruin condition | Welfare policy · Long-run welfare under ruin condition | long-term social welfare | Helped |
| Evaluate policies over a long horizon rather than a short horizon | Rawlsian versus utilitarian policies · SIPP simulations | social welfare | Mixed |
| Use random allocation instead of utilitarian allocation under survival conditions | Random policy · Long-run welfare under survival condition | long-term social welfare | Helped |
| Increase the budget M for random allocation | Random policy · SIPP simulations | convergence rate | Helped |
| Decrease b, strengthening the poor-get-poorer decay effect | Rawlsian min-U versus utilitarian max-U · Heterogeneous-population simulations | long-term social welfare | Helped |
| Change decay monotonicity from decreasing to constant | Rawlsian versus utilitarian policies · Survival condition, uniform bounds | long-term social welfare | No clear effect |
| Change decay monotonicity from decreasing to increasing | Rawlsian versus utilitarian policies · Survival condition, uniform bounds | long-term social welfare | Hurt |
| Change return and decay functions to constant | Rawlsian versus utilitarian policies · Survival condition, uniform bounds | long-term social welfare | No clear effect |
Surprising findings
- Rawlsian policies, despite being suboptimal in the short run, outperform utilitarian policies in the long run under the survival condition.
Limitations
- Results rely on uniform boundedness of return and decay bounds across individuals; heterogeneous cases are only studied by simulation.
- Simulations use simplified Gaussian-noise dynamics and welfare proxies from SIPP income bins.
Experimental setup
Datasets: Survey of Income and Program Participation (SIPP)
Metrics: long-term average social welfare (asymptotic growth rate) finite-horizon average welfare growth rate
Baselines: max-U (utilitarian) max-fg max-f max-g min-U (Rawlsian) random policy
Builds on: attention allocation model (staying with a winner / putting out fires) Lundberg's inequality
Closest papers by meaning
- Policy Learning for Balancing Short-Term and Long-Term Rewards (ICML 2024 · similarity 0.85)
- The Relative Value of Prediction in Algorithmic Decision Making (ICML 2024 · similarity 0.84)
- On the Meta-Design of Allocation Problems (NeurIPS 2026 · similarity 0.84)
- Navigating the Social Welfare Frontier: Portfolios for Multi-objective Reinforcement Learning (ICML 2025 · similarity 0.84)
- The Hidden Cost of Waiting for Accurate Predictions (ICLR 2025 · similarity 0.83)
- A Causal Lens for Learning Long-term Fair Policies (ICLR 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.